Quantitative morphology as a marker of cellular and organismal state

最近在模式识别方面的研究表明，计算机可以与人类专家的图像分类和模式分析相提并论，甚至超过它们。现代成像系统在空间和光谱分辨率以及动态范围方面远远超过人眼，因此有可能使基于机器的图像模式分析系统超过人类执行这些任务的能力。 我们开发并表征的模式分析系统称为WND-Charm。该方法基于使用标准特征提取算法以及我们自己开发的算法从每幅图像中提取超过2000个图像内容描述符。我们介绍的一项关键创新是不仅从原始图像中提取图像内容，而且还从图像变换(如傅立叶变换和小波变换)中提取图像内容。每个描述符基于其在训练中使用的图像集之间的区分能力而被分配分数。因此，任何图像都可以被绘制为该高维加权特征空间中的点。表示训练图像的点集被用来对表示每个训练类中存在的变化的概率密度函数进行建模。因此，可以根据图像的加权图像描述符和每个类别的概率密度函数来确定给定测试图像属于每个训练类别的概率--边缘概率。 这种分类机制的一个关键特性是，边际概率可以解释为图像的相似性，从而产生相似性的定量度量，而不仅仅是定性分类。今年投入了大量的努力来描述这些定量相似性测量的特征。这方面的一个自然测试案例是与年龄相关的形态变化。我们能够证明，从图像计算的形态年龄与已知的时间年龄有很好的相关性。我们的分类系统的通用性使我们能够研究线虫组织的衰老，通过差示干涉对比(DIC)成像以及苏木精/伊红(H+E)染色的小鼠组织切片。量化生理年龄的能力使我们能够更详细地描述衰老过程，并导致我们发现与年龄相关的形态变化不是连续的，而是通过不同的形态状态进行的。 我们继续开发我们的高密度RNAi筛选技术，我们已经开始对来自小规模试点屏幕的图像数据进行表征。量化图像相似性的能力使我们能够证明，敲除已知具有强烈遗传或物理交互作用的基因会导致高度相似的表型。将其扩展到全基因组筛选，将允许生成表型相似性网络，并表征具有未知功能的基因。

Recent work in pattern recognition has demonstrated that computers can equal or even surpass image classification and pattern analysis by human experts. Modern imaging systems far exceed the human eye in spatial and spectral resolution as well as dynamic range, thus potentially allowing machine-based image pattern analysis systems to surpass a human's capacity for performing these tasks. The pattern analysis system we've developed and characterized is called WND-CHARM. The approach is based on extracting over 2,000 descriptors of image content from each image using both standard feature extraction algorithms as well as those we developed ourselves. A key innovation that we introduced was to extract image content not only from the original images, but also from image transforms such as Fourier and wavelet. Each descriptor is assigned a score based on its ability to discriminate between the sets of images used in training. Thus, any image can be plotted as a point in this high-dimensional weighted feature space. The set of points representing the training images are used to model a probability density function representing the variation present in each of the training classes. The probability that a given test image belongs to each of the training classes - the marginal probabilities - can thus be determined from the image's weighted image descriptors and the probability density functions for each class. A key property of this classification mechanism is that marginal probabilities can be interpreted as image similarities, thus producing quantitative measures of similarity rather than merely qualitative classifications. A substantial effort this year was devoted to characterizing these quantitative similarity measurements. A natural test case for this is age-related morphological change. We were able to demonstrate that a morphological age calculated from images correlates well with known chronological age. The generality of our classification system has allowed us to study aging in C. elegans tissues imaged with differential interference contrast (DIC) as well as mouse tissue sections stained with hematoxylin/eosin (H+E). The ability to quantify physiological age has allowed us to characterize the aging process in much greater detail, and lead us to the discovery that age-related morphological change is not continuous, but progresses through distinct morphological states. We have continued to develop our high-density RNAi screening technology, and we have begun to characterize image data from small-scale pilot screens. The ability to quantify image similarity has allowed us to demonstrate that knock-down of genes known to have strong genetic or physical interactions leads to highly similar phenotypes. Extending this to full-genome screens will allow generating phenotypic similarity networks and characterize genes with unknown functions.