RI: Small: Using Humans in the Loop to Collect High-quality Annotations from Images and Time-lapse Videos of Cells

细胞生物学家分析活细胞的显微图像序列，以了解细胞过程，例如，预防癌症或设计用于伤口愈合的生物材料。当科学家发现图像分析工作太费力而无法自己完成，而自动化方法太多、不可靠或难以使用时，研究进展就会放缓或受到影响。该项目开发图像分析软件，以共同利用人力和计算机资源，特别是在互联网上，创建高质量的图像解释。活细胞成像研究支持了解细胞过程和设计生物材料的基础研究。具有统计学意义的绩效评估工作可能会对计算机视觉的研究方法产生广泛影响。该研究探索如何利用人力和计算机资源，特别是在互联网上，解释细胞的图像和视频。首先，对用于分析活细胞图像的检测、分割和跟踪算法进行了广泛的基准研究。然后开发了计算机视觉方法来解决现有算法的主要挑战，例如，解释在延时显微镜视频中由于有丝分裂而出现的新细胞。设计了用于量化自动和手动获得的注释的质量以及多个注释之间的可变性的方法。建立了一种工具，以有效和高效地利用领域专家的专业知识，特别是细胞生物学家，来判断和选择分析细胞图像的自动化方法。设计并进行了互联网工作者分析图像的众包实验。这些外行工作人员的批注质量与领域专家和自动化方法的批注质量进行了比较。最后，开发了一个机器学习系统，该系统可以自动确定在有或没有人工参与的情况下可以准确地分析哪些类型的细胞图像或视频。

Sequences of microscopy images of live cells are analyzed by cell biologists to understand cellular processes, for example, to prevent cancer or design bio-materials for wound healing. Research progress is slowed or compromised when scientists find the image analysis efforts too labor-intensive to do themselves and the automation methods too numerous, unreliable, or difficult to use. The project develops image-analysis software to leverage human and computer resources together, in particular on the internet, to create high-quality image interpretations. Live-cell imaging studies support basic research to understand cellular processes and design biomaterials. The work on statistically significant performance evaluation can have broad impact on the research methodology in computer vision.The research explores how human and computer resources can be leveraged together, in particular on the internet, to interpret images and videos of cells. Initially, an expansive benchmark study of detection, segmentation, and tracking algorithms for analyzing images of live cells is conducted. Computer-vision approaches to address the major challenges for existing algorithms are then developed, for example, to interpret the emergence of new cells due to mitosis in time-lapse microscopy videos. Methods are designed for quantifying the quality of automatically and manually obtained annotations and the variability between multiple annotations. A tool is built to effectively and efficiently use the expertise of domain specialists, in particular, cell biologists, to judge and select automated methods that analyze cell images. Crowd-sourcing experiments in which internet workers analyze images are designed and conducted. The quality of these lay workers' annotations is compared to the quality of annotations by domain experts and automated methods. Finally, a machine learning system is developed that automatically determines which types of cell images or videos can be analyzed accurately with or without human involvement.