Open access deep learning solutions for imaging flow cytometry

成像流式细胞术（IFC），其中每个单个细胞的图像是在它流过细胞仪时获得的，可以在几分钟内测量数十万个单个细胞，将传统流式细胞术的高通量能力与单细胞成像相结合。IFC不仅测量总荧光强度，而且测量群体中每个细胞的荧光空间图像加上亮场和暗场图像。然后，研究人员能够从这些图像中测量数百个不同的参数，例如细胞大小，形状，粒度以及细胞中任何荧光生物标志物的位置和强度。通过IFC捕获的丰富信息使其成为使用高内容方法的理想候选者，以专业的多变量分析工具，然后可以将相似的细胞聚集在一起，识别大量群体中的稀有细胞，并寻找细胞之间的关系，如干细胞分化途径。以前，我们将传统的机器学习技术应用于成像流式细胞仪的数据输出，以执行这些类型的任务，并开发了软件解决方案，允许非专家用户将这些方法应用于他们自己的数据集。虽然这些方法已经证明非常成功，但用户仍然必须使用图像分析工具来测量特征，这需要图像分析的知识。然而，最近随着深度神经网络（DNN）的引入，人工智能领域发生了一场革命。这些DNN从人类大脑的工作中汲取灵感，其中许多层都通过一些基本的处理规则相互连接。多核图形处理单元的使用的最新进展提供了在合理的时间内训练这些复杂网络所需的计算能力，并且他们发现了对图像识别问题的传统机器学习方法的显着改进。这些算法的一个主要优点是，用户不需要像传统机器学习算法那样测量细胞特征或参数。因此，这些DNN提供了完美的解决方案，解决了在应用机器学习工具之前需要进行图像分析的问题。此外，DNN需要非常大量的单细胞图像来训练网络，这使得它们非常适合IFC数据的分析。在这个项目中，我们将开发开放获取的软件工具，允许非专家将IFC的输出文件直接输入到神经网络中。我们将使用各种现有的DNN，这些DNN已经针对现有的图像识别问题（例如字符识别）进行了优化，并将它们应用于一组IFC数据集。选择这些数据集不仅是为了代表一组特定的生物学问题，也是为了对机器学习算法提出特定的挑战。我们还将开发和训练我们自己的DNN，这些DNN将针对特定的数据集进行优化，并提供这些数据集。DNN的一个缺点是网络学习分类图像的方法对用户是隐藏的，为了解决这个问题，我们将研究网络的最后一层，这些层包含网络用于分类图像的最特定模式。我们将把这些模式与它们所识别的细胞联系起来，以提高未来网络设计的性能。最后，我们将使用这些功能来可视化细胞之间的关系，以了解细胞在种群中的进化。

Imaging flow cytometry (IFC), where an image of each individual cell is acquired as it flows through a cytometer can measure hundreds of thousands of individual cells in minutes, combining the high-throughput capabilities of conventional flow cytometry with single-cell imaging. IFC measures not only total fluorescence intensities but also the spatial image of the fluorescence plus bright-field and dark-field images of each cell in a population. The researcher is then able to measure hundreds of different parameters from these images, for example the cell size, shape, granularity and the position and intensity of any fluorescence biomarker in the cell. This rich information captured through IFC makes it an ideal candidate for the use of high-content approaches to specialist multivariate analysis tools which can then cluster similar cells together, identify rare cells in large populations and look for relationships between cells such as stem cell differentiation pathways. Previously we have applied traditional machine learning techniques to the data output from an Imaging Flow Cytometer to perform these types of tasks and developed software solutions to allow a not expert user to apply these methods to their own datasets. While the methods have proven very successful the user must still measure the features using image analysis tools which requires knowledge of image analysis. However recently there has been a revolution in the field of artificial intelligence with the introduction of deep neural networks (DNNs). These DNNs take inspiration from the working of the human brain with many layers all connected to each other with some basic processing rules. Recent advances in the use of mulit-core graphical processing units has provided the computational power required to train these complex networks in a reasonable amount of time and they have found significant improvements over traditional machine learning methods for image recognition problems. One major advantage of these algorithms is that the user is not required to measure cell features or parameters as required for traditional machine learning algorithms. Therefore these DNNs offer the perfect solution to the problem of removing the step of requiring image analysis before the application of machine learning tools. Also the DNN require very large numbers of single cell images for training the networks which make them ideally suited to the analysis of IFC data.In this project we will develop open access software tools to allow a non-expert to input the output file from an IFC directly into a neural network. We will use a variety of existing DNNs which have been optimised for existing image recognition problems (e.g. character recognition) and apply them to a set of IFC datasets. These datasets have been chosen not only to represent a set of specific biological problems but to also pose specific challenges to the machine learning algorithms. We will also develop and train our own DNNs which will be optimised to the specific datasets and make these available. One disadvantage of DNNs is the method the network learns to classify images is hidden from the user and to address this problem we will study the last layers of the network which contain the most specific patterns that the network uses for classifying images. We will correlate these patterns with the cells they identify in order to improve the performance of future network design. Finally we will use these features to visualise the relationships between cells in order to understand the evolution of the cells in a population.